Detailed Description
Embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. While certain embodiments of the present disclosure have been shown in the accompanying drawings, it is to be understood that the present disclosure may be embodied in various forms and should not be construed as limited to the embodiments set forth herein, but are provided to provide a more thorough and complete understanding of the present disclosure. It should be understood that the drawings and embodiments of the present disclosure are for illustration purposes only and are not intended to limit the scope of the present disclosure.
Aiming at the problems in the background technology, the application provides an electric automobile battery thermal management method based on predictive control. Fig. 1 is a flowchart of an electric vehicle battery thermal management method based on predictive control according to an embodiment of the application. Fig. 2 is a data flow diagram of a prediction control-based electric vehicle battery thermal management method according to an embodiment of the present application. As shown in fig. 1 and 2, the electric vehicle battery thermal management method based on predictive control according to the embodiment of the application comprises the steps of S1, obtaining a long-term information source, a short-term information source and current vehicle state data, S2, conducting long-term and short-term working condition fusion prediction based on the long-term information source, the short-term information source and the current vehicle state data to obtain a fusion working condition prediction profile and a prediction uncertainty level, S3, inputting the fusion working condition prediction profile and a current battery measurement state into a battery model to conduct battery state prediction to obtain a battery state prediction track, S4, conducting MPC optimization with uncertainty perception based on the battery state prediction track and the prediction uncertainty level to obtain an optimal BTM control sequence, S5, extracting a current BTM control instruction from the optimal BTM control sequence, and sending the current BTM control instruction to an underlying executor of a BTM system to obtain a real-time deviation index, and S6, determining whether to correct the current BTM control instruction based on comparison between the real-time deviation index and a preset threshold.
In step S1, a long-term information source, a short-term information source, and current vehicle state data are acquired. In particular, the long-term information sources include navigation route data, long-term weather forecast, user charging schedule and vehicle historical energy consumption data, and the short-term information sources include GPS positioning data, real-time vehicle speed/acceleration sensor data, V2X traffic information, short-term weather forecast/measured ambient temperature and analysis results of recent operation habits of a driver. It should be appreciated that long-term information sources such as navigation routes and long-term weather forecast provide macroscopic and trending operation scene predictions for the system, which is conducive to strategic thermal management planning, short-term information sources such as real-time GPS, V2X information and driver behavior capture immediate and local dynamic changes, can effectively correct long-term predictions, improve immediate accuracy of predictions, and current vehicle state data such as battery SoC and internal temperature provide accurate initial conditions and feedback references for all predictions and controls. The comprehensive utilization of the multi-dimensional information aims to solve the problems of large prediction deviation and poor adaptability caused by the fact that only a single information source is relied on or long-term information is not fully fused, and a more reliable model prediction control optimization strategy with uncertainty perception is formulated for the follow-up accurate prediction of the battery state.
Specifically, the step S1 is obtained by firstly starting the acquisition of the long-term information source through a vehicle-mounted human-computer interaction interface or a preset. For example, when a user enters a destination in a navigation system, such as a city center library, the system obtains detailed navigation route data from the navigation module, the data including not only total mileage and estimated total time consumption, but also a series of road segment characteristics, such as 5 km in the future as city express way, speed limit 80km/h, and then turns into 3 km urban congestion road segments, estimated average vehicle speed 20km/h, and possibly altitude profile information of key road segments. At the same time, the vehicle connects to the internet weather service through its communication module, downloads a long-term weather forecast for several hours in the future, for example, 6 hours in the future, and obtains an ambient temperature profile, such as one temperature forecast and solar radiation intensity level per hour, along the way or in the target area. If the user sets a charging schedule by vehicle setting or mobile phone App, for example, 22:30 tonight starts charging at home charging post, the target SOC is 90%, and the user charging schedule information is recorded by the system. The vehicle may also retrieve vehicle historical energy consumption data associated with the current driver or current route from its internal storage or cloud database, for example, an average energy consumption of 14.2 kWh/100km based on the current driver's historical driving style and similar historical trip for the planned route.
In parallel, the short-term information source and the current vehicle state data are continuously acquired at high frequencies. The vehicle-mounted GPS positioning module provides the current accurate longitude and latitude coordinates of the vehicle at a preset frequency, such as 1 Hz. The CAN bus of the vehicle broadcasts real-time vehicle speed such as an instantaneous vehicle speed of 45km/h and vehicle acceleration such as a current acceleration of 0.5m/s2 from a wheel speed sensor and an acceleration sensor in real time. If the vehicle is equipped with a V2X communication unit it will receive real-time traffic information from surrounding infrastructure or vehicles, for example congestion occurring 20 seconds in front of a 800 meter traffic light green light or 1 km in front, average vehicle speed 10km/h. Short weather predictions, such as predictions of ambient temperature within 30 minutes of the future, may be obtained by fusion of measured values of the on-board ambient temperature sensor with shorter time-lapse network weather data. The recent operation habit analysis result of the driver can judge that the current driving style is stable or aggressive by analyzing the accelerator opening, the brake frequency and the intensity of the driver in the past minutes. The current vehicle state data is directly obtained from a Battery Management System (BMS), and mainly includes the current state of charge (SOC) of the battery, for example, 68%, and real-time temperature distribution data measured by a plurality of temperature sensors inside the battery pack, for example, the current average battery temperature is 29.5 ℃. All of these real-time acquired long-term information sources, short-term information sources, and current vehicle state data are processed by the system integration as inputs to the subsequent modules.
In step S2, based on the long-term information source, the short-term information source, and the current vehicle state data, a long-term and short-term condition fusion prediction is performed to obtain a fusion condition prediction profile and a prediction uncertainty level. Accordingly, considering that the traditional prediction method often depends on information of a single time scale, long-distance planning of strategic and immediate adjustment of tacticity are difficult to achieve, so that the prediction accuracy is low, and the adaptability is poor especially in complex and changeable real driving environments. By fusing the macroscopic trend provided by the long-term information source and the instant correction brought by the short-term information source and combining the current vehicle state, a fusion working condition prediction profile which is closer to the actual situation in the future can be generated. More importantly, the level of uncertainty of the overall prediction can be quantified by analyzing the inherent uncertainties of the individual information sources and the deviations between them. The method not only improves the robustness of prediction, but also enables the follow-up model prediction control to actively consider the uncertainty in the optimization process, thereby developing a battery thermal management strategy which is better and safer to perform under actual disturbance, and effectively solving the problems of inaccurate prediction and lack of uncertainty perception in the background art.
Specifically, in one possible embodiment, fig. 3 is a flowchart of step S2 in a prediction control-based electric vehicle battery thermal management method according to an embodiment of the present application. As shown in FIG. 3, the step S2 of performing long-term and short-term condition fusion prediction to obtain a fusion condition prediction profile and a prediction uncertainty level based on the long-term information source, the short-term information source and the current vehicle state data comprises the steps of inputting the long-term information source into a long-term prediction module to obtain a long-term power demand profile, a long-term environment temperature profile, a solar radiation profile and a long-term charging event sequence S22 of inputting the short-term information source and the current vehicle state data into a short-term prediction module to obtain a short-term power demand profile and a short-term environment temperature profile S23 of updating the long-term power demand profile based on the short-term power demand profile to obtain an updated long-term power demand profile S24 of updating the long-term environment temperature profile based on the short-term environment temperature profile, and integrating the updated long-term environment temperature profile, the solar radiation profile and the long-term charging event sequence to obtain the fusion condition prediction profile S25.
It should be appreciated that in order to translate the raw, macroscopic long-term planning information into a quantified future load and environmental condition prediction that is of direct instructive significance to the battery thermal management system. The application predicts the main energy consumption and corresponding heat generation condition in the future driving process by inputting the long-term information source into the long-term prediction module, namely, the long-term power demand profile, the long-term environment temperature profile and the solar radiation profile define the influence of the external environment on the heat exchange condition of the battery in the vehicle running period, and the long-term charging event sequence indicates the concentrated heat load possibly brought by the charging behavior. These structured prediction profiles are the basis for performing prospective energy optimization and thermal management strategy formulation, so that the system can pre-judge the risk of overheat or supercooling of the battery which may occur in the future, thereby effectively solving the problem that the conventional control strategy mentioned in the background art lacks prospective.
Specifically, step S21 processes the generation of a long-term power demand profile, which first relies on a fine analysis of navigation route data, and extracts therefrom characteristics of each road segment in future routes, such as road segment type, e.g., high speed, urban area, mountain road, average sectional vehicle speed, e.g., estimated 80km/h for urban express road segments, estimated 20km/h for urban congestion road segments, gradient information, e.g., 2% up-slope or-1% down-slope, and estimated travel duration for each road segment. These features are then input into a pre-trained power prediction model. The model is a deep neural network, which comprises an embedded layer for processing category characteristics such as road section type, and is connected with two layers of LSTM (long short term memory network) units, each layer is provided with 64 hidden units for capturing time sequence dependency of driving conditions, and finally a predicted power value is output through a full connection layer. The input feature vectors of the neural network include normalized average vehicle speed, gradient value, type of road section of the single heat code, and estimated fixed power consumption of auxiliary equipment such as air conditioner (for example, preset to 1.2 kw when the cooling is started in summer). The model is trained offline through a large number of historical energy consumption data of the vehicle including various driving conditions and their corresponding actual power consumption. After training, the feature sequence analyzed by the current planned route is input into the model, and a long-term power requirement profile which covers the future, for example, 60 to 120 minutes and has a time resolution of, for example, one data point per minute can be obtained.
The generation of the long-term ambient temperature profile and the solar radiation profile is based primarily on the obtained long-term weather forecast. The module extracts from the weather forecast hours in the future, for example, hour-by-hour ambient temperature predictions and solar radiation intensity predictions in watts per square meter along a predetermined route or target area within 6 hours in the future. To obtain a continuous profile matching the power demand profile time resolution, such as one data point every 15 minutes, the module will smooth the original hour-by-hour predicted data points with a cubic spline interpolation algorithm to form a time-series long-term ambient temperature profile and long-term solar radiation profile, respectively.
The generation of the long-term charging event sequence then accurately resolves the user charging schedule. For example, the user has set that this tonight 22:30 starts charging at home charging peg, with a target SOC of 90%. A preset rated charge power of the household charging pile, for example 7 kw, is first obtained. Then, the amount of charge required to reach the target SOC90% is calculated in conjunction with the current state of charge of the battery, for example, 35% obtained from vehicle state data and the rated total capacity of the battery, for example, preset to 60 kwh, and the preset average charge efficiency, for example, 90% is taken into consideration. The required charge duration was estimated from this, (90% -35%) 60 kwh/(7 kwh 90%) to be approximately equal to 5.2 hours. Thus, a time series will be generated in which the corresponding charging power value is set to 7 kw for a period of about 5.2 hours from 22:30, and the charging power value is set to 0 at all other points of time except this period. The time resolution of the sequence is consistent with the power demand profile, e.g., one data point per minute. Together, these precisely generated profiles provide key long-term predictive inputs for subsequent thermal management of the battery.
Accordingly, long-term predictions, while giving macroscopic trends, have difficulty coping with sudden traffic conditions, immediate changes in driver behavior, or rapid changes in microenvironment. Therefore, in order to finely capture recent dynamic changes of the vehicle through real-time and high-frequency data updating, thereby generating a short-term prediction result which is more accurate and quick in response than long-term prediction, a short-term prediction module predicts power requirements and environmental temperatures within seconds to minutes in the future by using information such as GPS positioning, real-time vehicle speed/acceleration, V2X information, short-term weather, driver habit analysis and the like, so as to obtain a short-term prediction profile.
Specifically, step S22 processes by generating a short-term power demand profile, the short-term prediction module receiving as input a short-term information source. These data are fed as input features into a pre-trained neural network model at its original acquisition frequency, e.g., GPS at 1Hz, vehicle speed acceleration at 0.1 seconds, or after appropriate processing. The model may be a Recurrent Neural Network (RNN) that includes an input layer, an LSTM long short term memory network or GRU gated recurrent unit layer (e.g., 32 to 64 neural units each configured to effectively capture dynamic dependencies between time sequences), and a fully connected output layer. The input layer receives the above-mentioned short-term source serialization data in a short past time window, for example, the last 5 to 10 seconds. The RNN model performs supervised learning training offline, and the training data set contains a large amount of high frequency sensor data in a real driving scenario and its corresponding actual vehicle power consumption within a short future period (e.g., 30 seconds to 90 seconds in the future, with a time resolution of, for example, 1 second). During training, the network weights are adjusted by a back propagation algorithm using, for example, root Mean Square Error (RMSE) as a loss function. After training, the model can predict the power demand profile in a short time in the future according to real-time input.
For the generation of the short-term environmental temperature profile, the short-term prediction module mainly fuses the actual measurement value of the vehicle-mounted environmental temperature sensor and the data of the network short-term weather forecast. For example, current on-board sensors measure an external ambient temperature of 29.5 ℃, while acquired short-term weather services forecast that the temperature will stabilize at that level for 30 minutes in the future. At this point, the module may employ a simple persistence model, i.e., predicting that the ambient temperature will remain substantially at 29.5 ℃ for 15 to 30 minutes in the future. Or if the short forecast shows a clear trend of temperature change, e.g. the temperature will drop 0.5 ℃ after 15 minutes, a linear interpolation or a simple autoregressive model, e.g. parameters p, d, q can be preset to (1, 1), using the measured temperature sequence of the last few minutes and the forecast trend to generate a short term ambient temperature profile for a short time, e.g. 15 to 30 minutes, with a higher time resolution, e.g. one data point per minute. The order of the predetermined parameters such as the AR model may be determined based on historical data analysis. The final output short-term power demand profile and short-term ambient temperature profile will be used for subsequent fusion and fine tuning of the long-term prediction profile.
It should be appreciated that long-term power demand profiles, while based on macro planning (e.g., navigation routes), have a large predicted time span, and it is difficult to fully capture all dynamic changes during actual driving, such as sudden traffic congestion, immediate adjustment of driver behavior, or unexpected road condition changes. Therefore, the long-term power demand profile is updated to effectively correct the possible deviation of long-term prediction in the initial stage, so that the whole prediction profile is closer to the actual impending working condition, more accurate input is provided for the subsequent model prediction control, and the effectiveness of a battery thermal management strategy and the energy efficiency of a system are improved.
Specifically, in one possible embodiment, step S23, updating the long-term power demand profile based on the short-term power demand profile to obtain an updated long-term power demand profile includes replacing the short-term power demand profile with a corresponding portion of the long-term power demand profile to obtain the updated long-term power demand profile. Accordingly, long-term power demand profiles, while providing a trend for longer periods of time in the future, have a relatively late response to immediate changes. The short-term power demand profile is generated based on high-frequency real-time data, and can more accurately reflect the dynamic demands of the vehicle in a very short time in the future. By directly replacing the short-term prediction with the high precision and high resolution in the initial section of the long-term prediction, the MPC controller can ensure that the prediction data which is firstly relied on is closest to the real situation when the MPC controller makes an optimization decision, so that the effectiveness and the robustness of the control action are improved, and the emergency in actual driving can be better dealt with, which is a key step for optimizing the thermal management efficiency of the battery.
Specifically, step S23 is processed by first obtaining a long-term power demand profile and a short-term power demand profile that were previously generated. The short-term power demand profile covers a short future time window, e.g., 30-90 seconds in the future, and has a high temporal resolution, e.g., one data point per second. The long-term power demand profile covers a longer time span, e.g., 15 minutes into the future (MPC predicted time domain Np) or even longer, which may have a lower time resolution, e.g., one data point per minute. The update procedure first determines the length of time covered by the short-term power demand profile, e.g., ns minutes. Then, in the long-term power demand profile, starting from the current time, an initial portion corresponding to Ns minutes is truncated. This part will be completely replaced by the short-term power demand profile. In particular, the short-term power demand profile may need to be processed to match the time resolution of the subsequent portion of the long-term power demand profile, or smoothed at the splice, prior to replacement. For example, if the short-term power demand profile is one data point per second and the long-term power demand profile is one data point per minute, the short-term power demand profile may be averaged or sampled over the time points per minute to obtain a resolution on the order of minutes, and replaced. The final updated long-term power demand profile is formed with the initial Ns minutes part being short-term prediction with high accuracy and the subsequent parts being the remainder of the original long-term prediction.
Further, based on the deviation between the short-term power demand profile and the corresponding portion in the long-term power demand profile, the updated long-term power demand profile may be corrected based on the deviation, that is, instead of directly replacing the corresponding portion in the long-term power demand profile with the short-term power demand profile, the updated long-term power demand profile may be obtained by combining the short-term power demand profile and the corresponding portion in the long-term power demand profile, which may promote the accuracy of the updated long-term power demand profile, thereby promoting the accuracy of the fusion condition prediction profile, and, since the fusion condition prediction profile also takes into account the deviation between the short-term power demand profile and the corresponding portion in the long-term power demand profile, this may also further promote the consistency between the fusion condition prediction profile and the prediction uncertainty level based on long-term information. That is, here, the deviation between the short-term power demand profile and the corresponding part in the long-term power demand profile is essentially due to the non-stationary characteristic of the electric vehicle electric heating system power demand time series data, i.e. the statistical characteristic difference related to the power time scale distribution in the time sequence direction.
Based on this, in another possible embodiment, step S23, updating the long-term power demand profile based on the short-term power demand profile to obtain an updated long-term power demand profile includes:
Obtaining a time series of power demand deviations from the short-term power demand profile and the long-term power demand profile by obtaining a time series of power demand deviations from corresponding portions of the short-term power demand profile and the long-term power demand profile, i.e., first obtaining a time series of corresponding portions of the short-term power demand profile and the long-term power demand profileTime series of power demand deviations。
Constructing a deviation delay phase space of each power demand deviation value in the time sequence of the power demand deviation to obtain a time sequence of the power demand delay phase space deviation value, namely: Wherein, the method comprises the steps of,AndRespectively represent the firstAnd (b)The time point at which the time point is the same,The time difference is indicated as such,AndRespectively in time series of power demand deviationsAndThe corresponding deviation of the power demand is used,Is the time series of power demand delay phase space deviation valuesThe individual power demand delay phase space offset values, that is, in the case of uniformly dividing the time series, may be subjected to offset delay phase calculation of time series scale decomposition, thereby constructing an offset delay phase space in the case of scale decomposition under the time series.
Performing conformal mapping calculation on each power demand delay phase space deviation value in the time sequence of the power demand delay phase space deviation values to obtain a time sequence of power demand deviation conformal mapping values, namely: Wherein, the method comprises the steps of,Is the time series of power demand delay phase space deviation valuesThe individual power requirements delay phase space deviation values,Is a logarithmic function value based on a natural constant e,Is the time series of the conformal mapping value of the power demand deviationThe power demand bias conformal mapping value, namely, the phase space conformal mapping is used for eliminating distortion among different time sequence phases, so that the bias uniform mapping under scale decomposition is ensured, namely, the decomposition scale coupling mapping of the time sequence fractal characteristics of the bias in the generalized phase space is realized.
Dynamically compensating and optimizing the time sequence of the power demand deviation based on the time sequence of the power demand deviation conformal mapping value and the time sequence of the power demand delay phase space deviation value to obtain the time sequence of the power demand compensation deviation, namely: Wherein, the method comprises the steps of,Coefficients for the weight of correction terms, e.g.Of course, this is merely an example, which may be determined from a priori knowledge, by experimental calibration or optimization algorithm adjustment,Is the time series of power demand compensation deviationsCompensating for deviations in power demand, i.eThe compensated power demand compensates for the bias, i.e., the decomposition scale timing distribution redundancy in the timing dimension is performed by introducing the initial bias value into the conformal mapping phase space, and the timing dependence across the bias delay phase space is constructed.
Updating the long-term power demand profile based on the time series of power demand compensation deviations to obtain the updated long-term power demand profile. In this way, compensating for deviations based on optimized power requirementsTo obtain an updated long-term power demand profile, for example, taking a half deviation position between the short-term power demand profile and a corresponding part in the long-term power demand profile as the updated long-term power demand profile, the accuracy of the fusion condition prediction profile and the consistency between the fusion condition prediction profile and the prediction uncertainty level based on long-short term information can be improved.
Likewise, the inherent uncertainty and low temporal and spatial resolution of long-term weather forecast makes it difficult to accurately capture the local and instantaneous ambient temperature changes actually encountered during vehicle travel. For example, long-term forecasts may not reflect micro-environmental temperature fluctuations caused by short-term vehicle entry into tunnels, underbridge shadow areas, or urban heat island effects, among others. Therefore, the deviation of the long-term prediction in the initial stage can be effectively corrected by updating the long-term profile, so that the updated environment temperature prediction is closer to the real situation about to be experienced by the vehicle, more reliable environment input is provided for the follow-up accurate battery state simulation and the optimization of the thermal management control strategy, and the overall control performance and the energy efficiency are improved.
Specifically, in one possible embodiment, step S24, updating the long-term ambient temperature profile based on the short-term ambient temperature profile to obtain an updated long-term ambient temperature profile includes replacing the short-term ambient temperature profile with a corresponding portion of the long-term ambient temperature profile to obtain the updated long-term ambient temperature profile. It should be appreciated that long-term ambient temperature profiles, while providing a trend of ambient temperatures for a longer period of time in the future, have a low temporal resolution and are difficult to capture local, rapid ambient temperature changes caused by microclimate, occlusion, etc. factors in the actual path of travel of the vehicle. The short-term environment temperature profile is based on real-time data of the vehicle-mounted sensor and short-term weather information updated by high frequency, and can more accurately reflect the environment temperature actually experienced by the vehicle in the future minutes to tens of minutes. By directly replacing the initial counterpart of the long-term prediction with this high-precision short-term prediction, it is ensured that the closest actual ambient temperature input is obtained during the near-term prediction period, thereby avoiding control deviation or energy consumption waste due to inaccurate ambient temperature prediction. In particular, the processing of this step is the same as that described above for the first embodiment of updating the long-term power demand profile.
Specifically, in another possible embodiment, step S24 may also perform the same process to obtain the updated long-term environment temperature profile by performing offset compensation on the short-term power demand profile and the long-term power demand profile to obtain the updated long-term power demand profile.
Accordingly, the temperature change of the battery is a result of the combined action of various factors, including power demand during driving (heat generation), ambient temperature (convection heat exchange), solar radiation (radiant heat absorption), and power input during charging (charge heat generation). Considering any one factor alone can not accurately predict the temperature rise of the battery, so that the key external and internal disturbances affecting the thermal state of the battery are required to be integrated to form a comprehensive fusion working condition prediction profile.
Specifically, step S25 processes by first ensuring that all of the input profiles, i.e., the updated long-term power demand profile, the updated long-term ambient temperature profile, the solar radiation profile, and the long-term charge event sequence, have a uniform predicted time range and time resolution. For example, the predicted future time range is set to the next Np time steps, e.g., np may be set to 30 steps, each step Δt being 1 minute, the total predicted time period being 30 minutes. If the temporal resolution of each profile at the time of generation is different, a resampling process is required first. For example, if the solar radiation profile is given with one data point every 15 minutes, while the other profiles are given with one data point every 1 minute, then interpolation (e.g., linear interpolation or constant hold interpolation) of the solar radiation profile is required to match the 1 minute resolution. Subsequently, for each discrete point in time k in the predicted time range, from k=0 to k=np-1, the respective profile is taken at that point in timeThe predictors at that point are combined. Specifically, the fusion condition prediction profile is at a point in timeThe entry for will include the entry from at the point in timeThe power demand value from the updated long-term power demand profile, the ambient temperature value from the updated long-term ambient temperature profile, the solar radiation intensity value from the solar radiation profile, and the charge power value from the long-term charge event sequence. In this way, a multi-dimensional time series is formed in which each time point is associated with a complete set of operating parameters. For example, at 5 minutes (k=4), the fusion operating mode prediction profile may include { power demand 10 kw, ambient temperature 25 degrees celsius, solar radiation 400 watts per square meter, charging power 0 kw }. This structured fusion condition prediction profile will then serve as key input to the battery thermal model and the MPC controller.
Specifically, in one possible embodiment, based on the long-term information source, the short-term information source, and the current vehicle state data, long-term and short-term operating condition fusion predictions are made to obtain a fusion operating condition prediction profile and a prediction uncertainty level, including obtaining weather forecast accuracy and navigational traffic prediction accuracy, calculating a deviation between the short-term power demand profile and a corresponding portion of the long-term power demand profile, and generating a short-term correction uncertainty level based on the deviation, and calculating a weighted sum of the weather forecast accuracy, the navigational traffic prediction accuracy, and the short-term correction uncertainty level to obtain the prediction uncertainty level.
The specific processing is as follows, firstly, weather forecast accuracy and navigation traffic forecast accuracy are obtained. The weather forecast accuracy is set to a value between 0 and 1, such as 0.85, which can be obtained based on long-term statistical evaluation of the historical forecast data of the weather service provider, such as temperature, radiation and actual observation, for example, if the historical data shows that the average absolute error of the 24-hour temperature forecast is less than a preset threshold, such as 85% of the probability of 1.5 ℃, the accuracy can be set to 0.85. Navigation traffic prediction accuracy, similarly, set to 0.75 for example, may be quantified by analyzing the deviation between the average speed or travel time predictions of the road segments provided by the navigation module history and the actual travel record of the vehicle, e.g., if the history data shows 75% the predicted travel time to actual travel time error is within a preset 10%.
Subsequently, a deviation between the short-term power demand profile and a corresponding portion of the long-term power demand profile is calculated, and a short-term correction uncertainty level is generated based on the deviation. Here, for example, the short-term power demand profile covers future Ns minutes such as ns=3 minutes. The portion of the long-term power demand profile corresponding to these Ns minutes is extracted. The two profiles are compared point by point for each time point, e.g. for a power value of one point every 10 seconds, and the average value of the absolute values of the differences is calculated, resulting in an average power deviation value, e.g. 3 kw. The short term correction uncertainty level is generated based on the average power deviation value, for example, it may be normalized, and if a significant upper limit of the preset power deviation is 10 kw, the short term correction uncertainty level may be calculated by dividing the average power deviation value by the upper limit of the preset power deviation, i.e., dividing 3 kw by 10 kw, to obtain 0.3. This value reflects the magnitude of the correction of the short-term real-time information to the long-term plan, the greater the correction, the greater the uncertainty contribution in this respect.
Finally, a weighted sum of the weather forecast accuracy, the navigational traffic forecast accuracy, and the short-term correction uncertainty level is calculated to obtain the forecast uncertainty level. The accuracy index is first converted into an uncertainty index, for example, the weather forecast uncertainty is 1 minus the weather forecast accuracy (i.e., 1 minus 0.85 is 0.15), and the navigation traffic forecast uncertainty is 1 minus the navigation traffic forecast accuracy (i.e., 1 minus 0.75 is 0.25). The three uncertainty sources weather forecast uncertainty, navigation traffic prediction uncertainty, short term correction uncertainty levels are then assigned preset weighting factors, for example, the weights can be respectively set to w1=0.2, w2=0.3, w3=0.5, the sum of the weighting factors is 1, and the specific values can be predetermined according to sensitivity analysis or expert experience of different factors on battery thermal management influences. The prediction uncertainty level is finally calculated from the sum of the products of the three uncertainty components and their corresponding weights, i.e., (0.15 x 0.2) + (0.25 x 0.3) + (0.3 x 0.5), resulting in a prediction uncertainty level of 0.255.
In step S3, the fusion condition prediction profile and the current battery measurement state are input into a battery model to perform battery state prediction so as to obtain a battery state prediction track. It should be appreciated that the performance, life and safety of an electric vehicle battery are highly dependent on its operating state, particularly temperature. The goal of thermal management is to actively maintain the battery within an optimal operating range. In order to achieve such active control, it is necessary to know in advance how the battery state will change under expected future driving and environmental conditions. The fusion condition prediction profile provides detailed information of future external excitation, and the current battery measurement state provides accurate initial conditions for prediction. The battery model can calculate and output a state sequence of the battery, namely a battery state prediction track, which changes with time in a future prediction time domain according to the inputs and following physical and chemical rules. The track is the basis for the following model predictive control to formulate the optimal thermal management strategy.
Specifically, in one possible embodiment, the method for predicting the battery state by inputting the fusion working condition prediction profile and the current battery measurement state into a battery model to predict the battery state comprises the steps of using the current battery measurement state as an initial condition, and using the battery model to forward simulate and calculate the evolution of the future battery state under the fusion working condition prediction profile to obtain the battery state prediction track.
The specific process is as follows, firstly, a calibrated electric-thermal coupling battery model is adopted. The model mainly comprises an electrical submodel and a thermal submodel. The electrical submodel, such as a second-order RC equivalent circuit model, has core parameters including ohmic internal resistance, polarized internal resistance and polarized capacitance, which are functions of battery state of charge (SOC) and temperature, and open circuit voltage, which are prestored in the form of a lookup table, and data of the electrical submodel originate from laboratory testing of battery cells or modules under different working conditions. The thermal sub-model is built using a lumped parameter method to describe the average temperature change of the battery pack as a whole, or a finer distribution parameter model is used to describe the internal temperature distribution. The key parameters include total heat capacity of the battery pack, for example, preset to 400 kilojoules per kelvin, which is determined based on the estimation of specific heat capacity, mass, etc. of the battery material, the convective heat transfer coefficient of the battery surface to the environment, for example, calculated according to a wind speed lookup table or by an empirical formula, can be set to 5 watts per square meter kelvin at rest, can reach 20 watts per square meter kelvin at running, the battery surface area is 1.5 square meters, for example, and the solar radiation absorption coefficient is 0.7, for example.
In simulation practice, the current battery measurement state obtained by the controller from the sensor is recorded as 0 in the predicted starting time, for example, the average temperature of the battery pack is 29.5 ℃, the highest cell temperature monitored in the pack is 30.8 ℃, and the current battery state of charge is 68%, which are initial calculation conditions of the battery model. Then, the simulation calculation sets the time step defined by the prediction profile according to the fusion condition, for example, the step to be 1 minute forward, and the total number of the forward prediction steps is 30, for example, the total number of the prediction steps corresponds to the prediction time domain of 30 minutes. At each discrete time step, advancing from 0 to the total number of predicted steps minus 1, first, the external input data corresponding to that time step, i.e., the power demand value at that time, e.g., 10 kilowatts discharged, the ambient temperature value, e.g., 25 degrees celsius, the solar radiation intensity value, e.g., 400 watts per square meter, and the charge power value, e.g., 0 kilowatts, is extracted from the fusion condition prediction profile. And secondly, calculating the actual working current value of the battery through an electronic model based on the power requirement or the charging power of the current time step and combining the state of charge and the average temperature of the battery at the moment. For example, the discharge current may be obtained by dividing the power demand value by the currently estimated battery terminal voltage, which itself is the product of the open circuit voltage minus the current value and the total internal resistance of the battery, which calculation may involve iterative solutions or rely on pre-stored detailed look-up tables. Then, based on the calculated current value and the total internal resistance of the battery in the current state, joule heat generated inside the battery is calculated. At the same time, it is also possible to take into account the reversible entropy heating (related to the current, temperature and the rate of change of the open circuit voltage with temperature) for the total heat production. The total heat production in this time step is obtained. And then, calculating heat exchange between the battery and the external environment by using a thermal sub-model according to the predicted environmental temperature value and the predicted solar radiation intensity value of the fusion working condition and combining the average surface temperature of the current battery. This includes the convective heat transfer calculated from the convective heat transfer coefficient, surface area and temperature difference, and the solar radiation heat absorption calculated from the solar radiation absorption coefficient, effective illuminated area and solar radiation intensity. Then, the total heat generation amount inside the battery and the net heat exchange amount with the environment (heat absorption and heat dissipation) are comprehensively considered, and the net increased heat inside the battery is obtained. Then, the variation of the average temperature of the battery in the time step is calculated according to the total heat capacity of the battery pack. For example, if the net heat input is 120 kilojoules over a time step and the total heat capacity is 400 kilojoules per kelvin, the temperature will rise by 0.3 degrees celsius. Thereby updating the average temperature of the battery at the beginning of the next time step. If a distributed parametric thermal model is used, the temperature of each node within the battery pack is similarly updated and the predicted highest cell temperature is recorded. Finally, the state of charge of the battery is updated according to the calculated operating current value and the time step. The new state of charge is equal to the current state of charge minus (or plus, for charging) the product of the current and the time step divided by the rated capacity of the battery (in ampere hours). And storing the updated average temperature, the highest single body temperature and the state of charge of the battery, which are obtained through calculation in the time step, as a data point of the predicted track of the state of the battery at the future time point. At the same time, these newly calculated state values become initial conditions for the next time step simulation calculation. The iterative calculation process is repeated for a total number of prediction step sizes, and finally a sequence which covers the whole prediction time domain and comprises the time evolution of the battery key state parameters, namely the battery state prediction track is generated.
In step S4, MPC optimization with uncertainty awareness is performed to obtain an optimal BTM control sequence based on the battery state predicted trajectory and the predicted uncertainty level. Accordingly, since conventional MPC optimization relies solely on nominal predictions, in practice control effects may be poor and even violate constraints such as battery temperature overrun due to prediction bias. The application introduces the prediction uncertainty level, and the MPC with uncertainty perception can explicitly consider the influence of the uncertainty in the optimization process, thereby improving the energy efficiency as much as possible while ensuring the safety and the service life of the battery.
Specifically, in one possible embodiment, performing MPC optimization with uncertainty perception based on the battery state prediction trajectory and the prediction uncertainty level to obtain an optimal BTM control sequence includes constructing an objective function, wherein the objective function comprises an energy consumption weight, a temperature deviation weight and a maximum temperature difference weight, the energy consumption weight, the temperature deviation weight and the maximum temperature difference weight are dynamically adjusted based on the prediction uncertainty level, acquiring constraint conditions, wherein the constraint conditions comprise state constraint, input constraint and uncertainty perception-constraint margin, and solving and optimizing the objective function based on an iterative algorithm in combination with the constraint conditions to obtain the optimal BTM control sequence.
The specific process is as follows, firstly, an objective function is constructed. In this scenario, the objective function is designed to minimize one weighted sum over the entire prediction horizon, e.g., np time steps, for a total duration of 30 minutes. I.e. objective function = (energy consumption term x energy consumption weight) + (temperature deviation sum term x temperature deviation weight) + (maximum temperature difference sum term x maximum temperature difference weight). Specifically, the weighted sum comprises several key parts, namely the first part is the estimated total energy consumption of the thermal management executing mechanism, such as the power consumption accumulated value of the cooling fan, the water pump and the refrigeration compressor in each time step, multiplied by the time step, and accumulated in the whole prediction time domain, wherein the weight is the energy consumption weight. The second part is the sum of squares of deviations between the battery temperature and its ideal target temperature, for example preset to 28 degrees celsius, which is determined according to the battery type and performance requirements, and the weight is the temperature deviation weight. The third part is the maximum temperature difference inside the battery pack, for example, the square of the difference between the predicted highest cell temperature and the average temperature, to ensure temperature uniformity, with the weight being the maximum temperature difference weight. The key is that the energy consumption weight, the temperature deviation weight and the maximum temperature difference weight are not fixed and are dynamically adjusted according to the input prediction uncertainty level. If the prediction uncertainty level is high, the temperature deviation weight and the maximum temperature difference weight tend to be increased to take a more conservative temperature control strategy to ensure safety, while the energy consumption weight may be appropriately reduced, because pursuing extreme energy conservation under high uncertainty may lead to a temperature overrun risk. The specific adjustment rule may be preset, for example, weight=reference weight (1+product of adjustment factor and prediction uncertainty level), and the adjustment factor is calibrated in advance through simulation or experiment.
Secondly, constraint conditions are acquired and set. These constraints ensure that model predictive control always honors physical and safety boundaries during optimization. The constraint conditions mainly include a state constraint, for example, the average temperature of the battery in the battery state prediction track must be maintained between a preset lower limit, such as 15 degrees celsius, and an upper limit, such as 35 degrees celsius, the highest single battery cell temperature must not exceed an absolute safety upper limit, such as 45 degrees celsius, and the battery state of charge should be within an allowable range. The input constraints, i.e. the control quantity of the battery thermal management actuator, such as the fan speed, the cooling power, cannot exceed its physical operational range, e.g. fan speed 0 to 5000 revolutions per minute, cooling power 0 to 5 kw, at each time step k. In particular, to characterize uncertainty perception, an uncertainty perception-constraint margin is introduced. This means that when setting the state constraint, a certain tightening will be performed according to the level of prediction uncertainty. For example, if the nominal maximum temperature constraint is 45 degrees celsius and the predicted uncertainty level is 0.255, then the upper constraint limit that is actually used for optimization may be set to 45 degrees celsius minus a margin value determined by the uncertainty level (e.g., margin equal to a sensitivity coefficient multiplied by the predicted uncertainty level multiplied by a temperature range reference, such as 0.1 multiplied by 0.255 multiplied by 10 degrees celsius, resulting in a margin of about 0.255 degrees celsius, then the actual constraint becomes 44.745 degrees celsius), thus leaving more margin for safety in the presence of greater uncertainty in the prediction.
Finally, combining the objective function and all constraint conditions, and solving the optimization problem based on an iterative algorithm. Because the model predictive control problem is a nonlinear and constrained multivariable optimization problem, mature numerical optimization algorithms such as Sequence Quadratic Programming (SQP), interior point method and the like are often adopted. The optimizer searches a group of battery thermal management control input sequences of Np time steps in the future prediction time domain, such as fan gear, refrigerating system power and the like of each time step, by using the latest battery state prediction track and prediction uncertainty level at the beginning of each control period through iterative calculation, so that the sequence can minimize the weighting objective function value constructed in the prior art on the premise of meeting all constraint conditions. The series of control actions obtained by solving are the optimal BTM control sequence.
In step S5, current BTM control instructions are extracted from the optimal BTM control sequence and sent to the underlying executor of the BTM system to obtain a real-time deviation index. It will be appreciated that the model predictive control calculates an optimal control sequence for the future time domain (e.g., np time steps) at each control cycle, but only executes the first control instruction in the sequence. Therefore, the control instruction corresponding to the current moment needs to be extracted from the whole optimization sequence for issuing. And sending the abstract control quantity to an underlying executor to convert the abstract control quantity into a key step of actual physical actions. The real-time deviation index reflects the idea of closed-loop control, and by monitoring the deviation between the actual execution effect and the expected value, the deviation information can be used for updating and correcting the battery model parameters, evaluating the control performance or serving as the input of the next round of prediction uncertainty evaluation, so that the accuracy and the robustness of the thermal management are continuously improved.
Specifically, in one possible embodiment, the current BTM control instruction is extracted from the optimal BTM control sequence and sent to an underlying executor of a BTM system to obtain a real-time deviation index, and the method comprises the steps of extracting the current BTM control instruction from the optimal BTM control sequence, inputting the current BTM control instruction into the underlying executor of the BTM system, obtaining a current real-time measured value, and calculating the difference between the current real-time measured value and the current state predicted in the last step to obtain the real-time deviation index.
The specific process is as follows, first, extracting a current BTM control instruction from the optimal BTM control sequence. After the model predictive control MPC in the previous step is optimally solved, an optimal BTM control sequence covering the future whole prediction time domain, for example, np time steps, where each time step Δt is 1 minute, and the total duration Np minutes is obtained. This sequence contains a proposed control action for each time step in the future, for example a sequence of Np elements, each of which may contain the target rotational speed of the cooling fan, the target power of the refrigeration compressor, the target power of the heater, etc. According to the rolling time domain characteristic of model predictive control, only the first element in the sequence is adopted as a control instruction to be executed at the current moment, namely the current control period. For example, if the optimal fan speed sequence is [2500 rpm, 2300 rpm, ], 1800 rpm ], then the current BTM control instruction extracted will have a fan speed fraction of 2500 rpm. Similarly, if the optimal cooling power sequence is [1.2 kw, 1.0 kw,..0.5 kw ], then the current cooling power command is 1.2 kw.
Next, executing the current BTM control instruction input to the BTM's underlying executor and obtaining a current real-time measurement. The current BTM control command, such as a fan speed of 2500 rpm and a cooling power of 1.2 kw, is converted into a signal (e.g., a voltage signal, a pulse width modulated PWM signal, or a message sent via an on-board network, such as a controller area network CAN bus) that CAN be recognized and executed by the underlying actuators and sent to the corresponding hardware units, such as a fan drive module and a compressor controller. These underlying executors will implement the received instructions. After the current control command is issued and acted for a period of time, and after the current control period is finished, the current real-time measured value is obtained through sensors of the battery management unit, such as a temperature sensor, a voltage sensor, a current sensor and other related sensors on the vehicle before the next control period is started. These measurements reflect the actual state of the battery after the control command has been executed. For example, the measured average temperature of the current battery pack is 30.5 degrees celsius, the highest cell temperature is 31.8 degrees celsius, and the current state of charge (SOC) of the battery is 67.2%.
Finally, the difference between the current real-time measured value and the current state predicted in the last step is calculated to obtain the real-time deviation index. The current state predicted in the last step refers to the predicted value of the battery model for the current battery state at the moment when the MPC optimization is performed in the last control period. For example, before the last control period, i.e., 1 minute, the MPC predicts that the average temperature of the battery pack at the current time should be 30.2 degrees celsius. The average temperature of 30.5 degrees celsius measured in real time is now compared with the predicted value of 30.2 degrees celsius for this last step. The difference is calculated as real-time deviation index average temperature= (30.5-30.2 degrees celsius) =0.3 degrees celsius. Similarly, the SOC deviation index, the maximum temperature deviation index, and the like can be calculated. The real-time deviation index quantifies the accuracy of model prediction and the actual effect of control execution, and is an important basis for subsequent model correction, uncertainty evaluation update or control strategy adjustment.
In step S6, it is determined whether to correct the current BTM control instruction based on a comparison between the real-time deviation index and a preset threshold. Accordingly, although model predictive control may be optimized according to predictions, complexity of actual conditions and imperfections of models may still result in deviations between predictions and actual states, i.e., real-time deviation indicators. If this deviation exceeds a preset threshold of a preset acceptable range, it indicates that the current control command may not be sufficient to effectively maintain the battery state within the desired range, or may result in excessive power consumption. By comparing and triggering the correction mechanism, the control command issued at present can be finely adjusted in quick response to the larger deviation, rather than being completely corrected by the next MPC optimization cycle. The method is helpful for correcting control deviation more timely, and preventing small deviation from accumulating into big problems, thereby improving the accuracy and safety of battery temperature control and the efficiency of overall heat management, avoiding unnecessary adjustment of tiny and normal fluctuation, and ensuring the control stability.
Specifically, in one possible embodiment, determining whether to correct the current BTM control instruction based on a comparison between the real-time deviation metric and a preset threshold includes fine tuning the current BTM control instruction based on fuzzy logic in response to the real-time deviation metric being greater than or equal to the preset threshold.
The method comprises the following steps of firstly, triggering the process under the condition that the real-time deviation index is larger than or equal to the preset threshold value. For example, if the average temperature deviation of the battery in the real-time deviation index is positive 0.8 degrees celsius (indicating that the actual temperature is 0.8 degrees celsius higher than the predicted value of the previous step), and the preset deviation threshold is positive 0.5 degrees celsius, the trimming mechanism is activated because the 0.8 degrees celsius is greater than 0.5 degrees celsius. This preset threshold, e.g., 0.5 degrees celsius, is determined based on an understanding of battery performance and safety requirements and extensive experimental data analysis, intended to distinguish normal, acceptable fluctuations from significant deviations requiring intervention.
Once triggered, fine-tuning based on fuzzy logic begins. The first step is blurring of the input variables. The real-time deviation index (e.g., the aforementioned average temperature deviation plus 0.8 degrees celsius) as an input is converted into a fuzzy linguistic variable. This requires predefining fuzzy sets describing the degree of deviation and their membership functions. For example, the average temperature deviation may be divided into fuzzy subsets of large deviation plus (DBP), medium deviation plus (DMP), small deviation plus (DSP), deviation Zero (DZ), small deviation minus (DSN), etc. For an input value of plus 0.8 degrees celsius, it may be positive with a membership of 0.7 for a deviation, while positive with a membership of 0.3 for a deviation. The form of membership functions (e.g., triangle, trapezoid, gaussian) and their parameters are set according to empirical or data-driven methods. The second step is fuzzy rule reasoning. The core is a set of preset IF-THEN fuzzy rule base. These rules map the blurred input deviation state to the amount of adjustment of the blurred output control instructions. For example, one rule may be that IF average temperature deviation is deviates more by plus THEN fan speed adjustment is increased more AND cooling power adjustment is increased more. Another rule may be that the positive THEN fan speed adjustment is increases by a medium AND the cooling power adjustment is increases by a medium in the IF average temperature deviation is deviation. These rules are activated to varying degrees depending on the membership of the input bias plus 0.8 degrees celsius to the bias plus. The third step is the aggregation and defuzzification of the fuzzy outputs. Multiple rules may be activated simultaneously, the outputs of which (e.g., fan speed adjustment is increased more and fan speed adjustment is increased more, etc.) need to be aggregated into a total fuzzy output. The total fuzzy output is then converted into an accurate, numerical control command adjustment by a defuzzification method (e.g., gravity, maximum membership). For example, the resulting fan speed adjustment may be 250 rpm and the cooling power adjustment may be 0.2 kw.
Finally, the calculated fine adjustment is applied to the current BTM control instruction. If the original current fan speed command is 2500 rpm and the cooling power command is 1.2 kw, the fine-tuned command is changed to a fan speed 2750 rpm (2500 rpm plus 250 rpm) and a cooling power of 1.4 kw (1.2 kw plus 0.2 kw). This trimmed control instruction will be issued to the underlying executor instead of the original instruction.
In summary, the electric vehicle battery thermal management method based on predictive control according to the embodiment of the application is explained, which generates a fusion working condition prediction profile considering global and local characteristics by fusing a long-term information source pre-short-term information source and current vehicle state data, introduces deviation dynamic compensation and conformal mapping optimization, improves prediction precision, and overcomes the limitation of traditional single information source prediction. And secondly, based on the weighted evaluation of the weather forecast accuracy, traffic forecast deviation and short-term correction uncertainty level, the prediction uncertainty level is quantized, and is embedded into the MPC optimization objective function weight dynamic adjustment and constraint margin design, so that the active perception and robust optimization of uncertainty are realized, and the excessive dependence of a controller on ideal prediction is avoided. In addition, by means of real-time deviation index monitoring and a fuzzy logic fine adjustment mechanism, control instructions are corrected online when the deviation between the prediction and the actual state exceeds a threshold value, and the self-adaption capability of the system to disturbance is enhanced. According to the scheme, through multi-source information fusion prediction, uncertainty perception optimization and closed-loop feedback correction, the prospective, robustness and instantaneity of a battery thermal management system are remarkably improved, and the problems of inaccurate prediction and poor immunity of a traditional method are solved.
Fig. 4 is a block diagram of an electric vehicle battery thermal management device based on predictive control according to an embodiment of the application. As shown in fig. 4, the electric vehicle battery thermal management device 100 based on prediction control according to an embodiment of the present application includes an information data acquisition module 110 configured to acquire long-term information source, short-term information source and current vehicle state data, a fusion condition prediction module 120 configured to perform fusion prediction of long-term and short-term conditions based on the long-term information source, the short-term information source and the current vehicle state data to obtain a fusion condition prediction profile and a prediction uncertainty level, a battery state prediction track generation module 130 configured to input the fusion condition prediction profile and a current battery measurement state into a battery model to perform battery state prediction to obtain a battery state prediction track, an information data acquisition module 140 configured to perform MPC optimization with uncertainty perception based on the battery state prediction track and the prediction uncertainty level to obtain an optimal BTM control sequence, a real-time deviation index calculation module 150 configured to extract a current BTM control instruction from the optimal BTM control sequence and send the current BTM control instruction to an underlying executor of a BTM system to obtain a real-time deviation index, and an instruction correction module 160 configured to determine whether to correct the current BTM control instruction based on a comparison between the real-time deviation index and a preset threshold.
As described above, the electric vehicle battery thermal management apparatus 100 based on predictive control according to the embodiment of the present application may be implemented in various wireless terminals, for example, a server or the like having an electric vehicle battery thermal management algorithm based on predictive control. In one possible implementation, the electric vehicle battery thermal management device 100 based on predictive control according to an embodiment of the present application may be integrated into a wireless terminal as one software module and/or hardware module. For example, the predictive control-based electric vehicle battery thermal management device 100 may be a software module in the operating system of the wireless terminal or may be an application developed for the wireless terminal, although the predictive control-based electric vehicle battery thermal management device 100 may be one of many hardware modules of the wireless terminal.
Alternatively, in another example, the predictive control-based electric vehicle battery thermal management apparatus 100 and the wireless terminal may be separate devices, and the predictive control-based electric vehicle battery thermal management apparatus 100 may be connected to the wireless terminal through a wired and/or wireless network and transmit the interactive information in a agreed data format. Here, it will be understood by those skilled in the art that the specific operations of the respective steps in the above-described predictive control-based electric vehicle battery thermal management apparatus have been described in detail in the above description of the predictive control-based electric vehicle battery thermal management method with reference to fig. 1 to 3, and thus, repetitive descriptions thereof will be omitted.